A client asked me a question last month that I suspect is being asked in a lot of boardrooms right now: “The AI agents can handle cases on their own now — so why are we still paying people to do it?” It is the wrong question, but it is wrong in an interesting way, and working through it is the best way I know to explain how autonomous service agents should actually be adopted.
What Is Actually Available Now
Microsoft shipped four AI agents for Dynamics 365 Customer Service to general availability back in October 2025: the Case Management Agent, the Customer Intent Agent, the Quality Evaluation Agent, and the Customer Knowledge Management Agent. The 2026 release wave 1 is deepening all four — richer telemetry, tighter Copilot integration, and crucially, support for semi-autonomous and fully autonomous workflows. In plain terms: the platform is now capable of receiving a case, understanding what the customer wants, resolving it, and grading its own work, without a human touching the record.
Capable, however, is not the same as advisable. The question is not whether the agent can handle a case. It is whether you should let it — and for which cases.
The Delegation Framework I Use
When I sit down with a service manager, I draw a simple two-by-two on the whiteboard: volume on one axis, cost of getting it wrong on the other.
High volume, low risk. Password resets, order status, return labels, “where is my invoice.” This is the quadrant where autonomous handling should start, and honestly, where it should have started years ago with plain workflows. The agent version is simply better at understanding messy human phrasing.
High volume, high risk. Billing disputes, cancellations, anything involving money leaving your account or a customer leaving your business. Semi-autonomous is the ceiling here: let the agent draft the response, gather the history, and propose a resolution — but a human clicks approve. The draft-plus-approval pattern captures most of the efficiency with almost none of the exposure.
Low volume, high risk. Regulatory complaints, legal threats, safety issues. Keep these fully human, and configure your intent routing so the agent’s first job is to recognize them and get out of the way fast. The Customer Intent Agent is arguably more valuable as a smoke detector than as a resolver.
Low volume, low risk. The long tail of odd, rare questions. Counterintuitively, I leave these with humans too — there is not enough volume to justify tuning, and rare cases are where agents hallucinate procedures that do not exist.
The Quality Evaluation Agent Is Your Trust Meter
The most underrated of the four agents is the one that does not talk to customers at all. The Quality Evaluation Agent grades interactions — human and AI alike — against your quality criteria. Used well, it becomes the instrument that tells you when to widen the autonomous lane.
The pattern I recommend: pick one case category from the high-volume, low-risk quadrant. Run it semi-autonomous for a month, with the Quality Evaluation Agent scoring every interaction. Compare the agent’s scores against your human baseline. When the agent consistently matches or beats the baseline — and only then — flip that one category to fully autonomous, and pick the next category. Expansion is earned with evidence, not enabled with a checkbox.
⚠️ Important: decide up front what your rollback trigger is. “If CSAT for autonomous cases drops below X, or escalation rate rises above Y, the category goes back to semi-autonomous.” Writing this down before go-live turns a future crisis meeting into a routine operational decision.
What This Means for the Team
Back to my client’s question — the reason it is the wrong question. Every service team I have worked with has a backlog of work they never get to: knowledge articles that are two versions out of date, proactive check-ins that never happen, root-cause analysis that stops at “resolved.” When the routine forty percent of case volume goes autonomous, that is where the human capacity should go. The teams that treat this as a headcount reduction exercise will run their knowledge base into the ground within a year — and ironically, the agents depend on that knowledge base more than the humans ever did.
Closing Thoughts
Autonomous case handling is no longer a preview feature you can politely ignore; it is a capability your competitors are already tuning. But adoption is a governance exercise, not a procurement one. Start in the safe quadrant, make the Quality Evaluation Agent your referee, write your rollback rules before you need them, and reinvest the freed-up hours in the knowledge layer that makes the whole system smarter. Delegation done well looks boring — and boring is exactly what you want your service operation to be.
